Skip to content

maksym-fedorchuk-quarks-tech/peeking_problem

Repository files navigation

peeking_problem

Here we going to do experiment for the Peeking Problem explanation

A brief definition of the Peeking Problem is that it arises when there is constant "peeking" at the test results before they reach statistical significance or the planned and predetermined duration of the test. As a result, this can lead to incorrect data interpretation.

Most of analytics do split tests (or are present during the process - it's how lucky they are), and generally, they know that a standard test (like the Student's t-test) doesn't answer the question, "Is the new payment screen better?" but rather answers, "What is the probability that the differences between our test and control groups are due to chance?". And if an analyst has agreed with a manager that the test will be accepted at a 95% significance level (meaning that they choose a 5% probability that the result is not random), the practice of peeking allows for the "accidental" acceptance of a test (which shouldn't have been accepted) with a much higher probability 🙂. The materials I've come across online suggest that, based on their experiments, this percentage can reach up to 20%. And this potentially means that every 5th rejected split-test rejected with no good reason 😣...

About

Peeking Problem experiment

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages